【作者】 沈美丽；

【导师】 陈殿仁； 孙俊喜；

【作者基本信息】 长春理工大学 ， 物理电子学， 2006， 博士

【摘要】 为了解决目前临床上通过显微镜、利用目视方法来定性地分析尿沉渣涂片，并根据尿沉渣有形成分的颜色和形态特征等进行人工分析、识别和计数等操作的弊端。本文设计出一套集病理分析、分类及诊断等功能为一体的尿沉渣有形成分分析系统，在实际应用中，该系统能快速、准确地对尿沉渣中的有形成分进行统计和分析，此系统既减轻医务人员工作压力，又提高了工作效率和诊断正确率，对尿沉渣病理诊断进行科学性分析具有重要意义。 本文从理论和实践两个方面研究了尿沉渣有形成分的优化检测方法，利用数字图像处理技术和模式识别理论对尿沉渣图像的分割和分类进行了深入研究。从尿沉渣图像本身特征出发，将多种处理算法有机的结合，通过计算和反复实验，提出了一套完整有效的对尿沉渣有形成分分析和判别的处理方案。 为了能够对尿沉渣有形成分进行准确的数据分析，关键是要对该种图像进行正确分割。本文采用分水岭变换对尿沉渣图像进行分割，但此算法容易造成图像的过分割，导致图像的轮廓线掩埋在杂乱的分水岭线中。为了避免这种过分割现象，首先在前人研究的基础上利用形态学变换算法对尿沉渣图像进行预处理，并对预处理后的图像进行标记提取和分水岭变换。这样，在去除噪声的同时，减少了过分割区域的数目，但有少部分仍然存在过分割。针对此缺陷，本文提出了将水平集方法和分水岭变换相融合的分割方法，实验证明，该方法不仅很好地解决了过分割问题，而且使尿沉渣有形成分能够准确、快速地分割出来，并且把在分割过程中产生的尖点、断裂等情况均融合为一体。对分割后的尿沉渣图像，利用形态特征参数对尿沉渣有形成分进行描述。采用轮廓跟踪、径向基神经网络和支持向量机等方法对尿沉渣有形成分分类识别，提出以轮廓跟踪方法计算得出的特征参数值作为支持向量机的输入训练样本，再经过反复训练和测试得到多个分类器，将被测图像输入到多分类器中进行分类，解决了尿沉渣有形成分的多分类问题。综上所述，作者认为本文的创造性工作如下。 (1) 本文在理论和实验研究的基础上，指出了利用分水岭算法对尿沉渣图像进行分割存在的过分割现象，为解决这一问题，作者提出了首先采用形态处理算法对尿沉渣图像进行预处理，以消除图像中的大部分噪声和无关杂质，再采用先标记图像后用分水岭分割图像的方法，并对分割后的图像通过区域合并处理得到最终的分割结果，此种方法在很大程度上解决了分水岭算法的过分割现象； (2) 尽管上述方法能在很大程度上解决过分割问题，但过分割现象还在某种程度上存在。作者根据过分割后小区域间像素点的灰度值比较接近的特点，提出了将水平集方法和分水岭算法相融合的尿沉渣图像分割方法。实验证明，这两种方法的结合，不仅减少了分割的盲目性，而且提高了分割的准确性； (3) 使用轮廓跟踪方法对大量的尿沉渣图像进行特征参数计算，得出了不同有更多还原

【Abstract】 In order to solve the default of using microscope observation and visual method determine the nature analysis urinary sediment smear and carrying manually analyze, identification and count according to color or form characteristic of the visible urinary sediment component, a set of urinary sediments visible component analysis system including pathology analysis, classification and diagnosis is proposed in this paper. In actual application, the system can make statistics and analyses rapidly and accurately which not only release the doctor workload, but improve work efficiency and diagnose correct rate and have important signification to analyzes urinary sediment pathology diagnosis.This paper researches optimization detection method of urinary sediment image visible component from theory and practice. It discusses deeply the segmentation and classification using digital image processing and pattern recognition theory. Through amount of computing and experiments, a set of complete effective processing scheme for urinary sediment visible component is proposed by combining multi-processing arithmetic from the characteristic of urinary sediment image.It is necessary to correctly segment image in order to analyze accurately urinary sediment visible component. This paper adopts watershed transform to segment those images, but this method easily products the over segmentation which results in the edge line buried in disorderly watershed line. In order to overcome that default, morphologic transform and signature extracting are used before watershed transform and used region merging method to remove or decrease the question of over segmentation. However, the over segmentation is still exist. In order to overcome this question, a kind of method which combines level set and watershed transform is proposed. It is proved by test that the proposed method not only resolve the question of over segmentation but has the ability accurately and speed to segment urinary sediment visible component. And then, using morphology feature parameter analyzes and describes segmented urinary sediment image. Using contour tracking, RBFNN and SVM three kind of different method classify urinary sediment visible component and using feature parameter value which is computed by contour tracking method input SVM as input train sample. Through reply train and test, the multi-classification is get which resolves the question of multi-classification of urinary sediment visible component. The author think that the creative work in this paper is as following:(1) Based on the theory and test research, the author has pointed out over segmentation question when using watershed algorithm segments sediment image. In order to solve that question, morphologic transform are used to pre-proposed image to remove更多还原